A model can contain the visual information needed to answer a question and still fail to use it. That gap between information being present and being selected is the paper’s central concern—and a more testable question than whether a model is conscious. The authors use “conscious” only for functional patterns associated with global workspace theory, not to claim subjective experience.
Key Findings
- The Multimodal Concept Lens reads concepts from intermediate vision and language states in a shared vocabulary space. This lets the researchers compare what is represented in the image pathway with what becomes accessible during question answering.
- Questions can change which objects are most accessible in the same scene. In experiments on visual question answering, the queried object became more prominent within the decoder’s estimated workspace than competing objects.
- Counterfactual interventions on concept representations could redirect answers, including in tasks requiring multiple reasoning steps. This suggests the readouts track information that can influence computation, rather than merely describing it.
- In controlled alignment experiments, concept accessibility improved before reliable query-dependent selection. The distinction offers a way to study different stages of vision-language alignment.
Who It’s For—and the Trade-off
A good fit for researchers studying multimodal interpretability, causal interventions, or functional analogies between neural networks and cognitive theories. The evidence is about internal representations and model behavior; it does not establish consciousness, and findings from selected models and experimental setups should not be treated as a general account of human cognition.